Vocal Separation Using Fourier Methods¶
This method relies on the fact that music is periodic (repeats in time) wheras vocals are relativly aperiodic. Fourier transforms transform data from time space to frequency space. The fourier transform for a simple sine wave is shown below
- By performing a short time fourier transform on a song, we can obtain a spectrogram showing the frequencies present in each small time window.
- By performing a second fourier transform on the spectrogram, we can produce a new image that shows peaks corresponding to periodic patterns spectrogram. Peaks along the x axis correspond to frequencies that repeat in time ie. music.
- By Separating the peaks from the rest of the data, we can separate the musical components from vocal components. Then by taking the separated components through the reverse process we can obtain two separated audio files.
Major limitation is that some vocal frequencies are periodic and some musical frequencies are aperiodic, so there is some leakage between files and perfect separation cant be obtained with this method. Need machine learning techniques for further separation.